mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 04:31:58 +00:00
ThearchyHelios
56ca53ad35
用量（第 1 波遗留）：查清楚了，丢的不是浏览器那边——问数据页一直读到 SSE 结束、usage 那一条收得到——而是服务端生成过程
（api/copilot_gen._iter_ops）读到 done、look、ask 那一行就关掉模型的流，流末尾那一段才报的用量（Anthropic 的输出 token、
OpenAI 兼容系的 stream_usage）一次都拿不到，每次建图都记成「估不出」，这一轮的花费多半是下限。改成交出那一行之后先把流读完
再关（_drain，内容丢掉）：照约定那一行就是最后一行，读完只多收用量那一段；模型在那一行之后还往下写超过 DRAIN_CHARS 字闲话的
当场关流、照实记成没报全（不为闲话付钱，也不当成报全了）。评测的作答（app/eval/answer.Driver）走同一个 answer_ops，
用量跟着报全，和问数据页同一个口径。test_usage_meter：原先断言「输出 token 为 0、记成估不出」的那条改成报全；另加 look 后、
闲话超限、评测和问数据页收到同一份用量三条；去掉修复回跑，三条都抓得住。providers/usage.py 的说明跟着改。
通知：notify_eval 的链接从定时任务的触发历史改成这次评测的详情页 /evals/runs/{id}（评测页这一波上线）；测试设上对外访问地址、
断言链接（原来的断言在没设地址时对空列表恒真）。
评测接口（只加不改，app/api/evals.py）：题集列表和详情带 latest（最近一次评测）、trend（同一变体最近几次评完的准确率，旧的
在前，换了模型、目录的不连成一条线）、schedules（目标是它的定时任务）；评测详情带 actions（停止、只重试没做完的、重新评分、
钉住 / 取消基线能不能做，和接口拒收的条件同一份）和没做完的项数；POST /suites/{id}/cases/preview 写 SQL 定期望的预览
（同期望的 SQL 探针，交回列和前 50 行，遮罩列照保存时那一道拒收、没过守卫的指明是哪一格，都不回 422）；GET /signals
「口径命中与重复度」的判据和分档、窗口趋势（数只出自 ledger/stats.compute_stats）、成员不少于 3 的问题类（只出自
chat/questions，评测运行不进）。为什么放服务端：界面层不放逻辑，统计、能不能做都由服务端给。tests/test_evals_ui_api.py。
374 lines
18 KiB
Python
374 lines
18 KiB
Python
"""用量计量（providers/usage.py；细案 01 做法 3）：助手建图（含挑表、接着写、自查修正）和复核的用量也记。
|
||
|
||
守的是：
|
||
- 按模型累计 tokens 和调用次数，金额按价格目录算；目录外的模型、回复没带用量（而目录价又不是 0）的调用记 unpriced_calls，
|
||
不当成 0 元；
|
||
- metered 只计生成过程自己的调用（含它在一步里起的任务）；调用方在两次 yield 之间做的事（发起运行）不计——运行的用量
|
||
另有 llm.end，包进来就重复了；嵌套时内层并进外层；
|
||
- answer_ops 末尾的 usage 记下这一轮建图的每一次调用（起草、自查修正各一次）；复核的结果带 usage；
|
||
- 读到 done / look / ask 那一行之后照样把模型的流读完再关（读完只为拿到流末尾才报的用量）：输出 token 报全了、金额估得出，
|
||
不再是下限；模型在那一行之后还往下写一大段闲话的，读到上限就当场关流、照实记成没报全；
|
||
- 评测的作答（app/eval/answer.Driver）记的助手用量就是问数据页收到的那条 usage：同一个口径。
|
||
|
||
模型都是剧本，不调真接口；价格用目录里登记过的模型 id，名字之外都是虚构的。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import json
|
||
from typing import Any
|
||
|
||
import pytest
|
||
from httpx import ASGITransport, AsyncClient
|
||
from langchain_core.language_models.chat_models import BaseChatModel
|
||
from langchain_core.messages import AIMessage, AIMessageChunk
|
||
from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
|
||
|
||
from app.main import app
|
||
from app.providers import usage
|
||
from app.providers.catalog import estimate_cost
|
||
|
||
PRICED = "claude-opus-5" # 价格目录里有
|
||
UNPRICED = "made-up-model-x" # 价格目录里没有
|
||
FREE = "mock-fast" # 目录价是 0
|
||
|
||
|
||
class Scripted(BaseChatModel):
|
||
"""真的 LangChain 模型：回调照常触发。每次调用吐 replies 里的下一份（用完重复最后一份），用量按 tokens 报；
|
||
report=False 时回复不带用量（有的网关流式不给）。"""
|
||
|
||
model: str = PRICED
|
||
replies: list[list[str]] = [["好的"]]
|
||
tokens: tuple[int, int] = (1000, 200)
|
||
report: bool = True
|
||
calls: int = 0
|
||
|
||
@property
|
||
def _llm_type(self) -> str:
|
||
return "scripted"
|
||
|
||
@property
|
||
def _identifying_params(self) -> dict[str, Any]:
|
||
return {"model": self.model}
|
||
|
||
def _next(self) -> list[str]:
|
||
lines = self.replies[min(self.calls, len(self.replies) - 1)]
|
||
self.calls += 1
|
||
return lines
|
||
|
||
def _usage(self) -> dict[str, int] | None:
|
||
if not self.report:
|
||
return None
|
||
inp, out = self.tokens
|
||
return {"input_tokens": inp, "output_tokens": out, "total_tokens": inp + out}
|
||
|
||
def _generate(self, messages, stop=None, run_manager=None, **kwargs) -> ChatResult:
|
||
message = AIMessage(content="\n".join(self._next()))
|
||
if (meta := self._usage()) is not None:
|
||
message.usage_metadata = meta
|
||
return ChatResult(generations=[ChatGeneration(message=message)])
|
||
|
||
async def _astream(self, messages, stop=None, run_manager=None, **kwargs):
|
||
"""像 Anthropic 那样报用量：第一段带输入 token，流末尾单独一段带输出 token。"""
|
||
lines = self._next()
|
||
inp, out = self.tokens
|
||
for i, line in enumerate(lines):
|
||
meta = {"input_tokens": inp, "output_tokens": 0, "total_tokens": inp} if self.report and i == 0 else None
|
||
yield ChatGenerationChunk(message=AIMessageChunk(content=line + "\n", **({"usage_metadata": meta} if meta
|
||
else {})))
|
||
if self.report:
|
||
yield ChatGenerationChunk(message=AIMessageChunk(content="", usage_metadata={
|
||
"input_tokens": 0, "output_tokens": out, "total_tokens": out}))
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 计量本身
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
async def test_metering_counts_by_model_and_prices_from_the_catalog():
|
||
m = usage.meter()
|
||
with usage.metering(m):
|
||
await Scripted(model=PRICED).ainvoke("你好")
|
||
await Scripted(model=PRICED).ainvoke("再来")
|
||
await Scripted(model=UNPRICED).ainvoke("你好")
|
||
await Scripted(model=PRICED + "-silent", report=False).ainvoke("你好")
|
||
await Scripted(model=FREE, report=False).ainvoke("你好")
|
||
got = m.summary()
|
||
assert got["by_model"][PRICED] == {"input_tokens": 2000, "output_tokens": 400, "calls": 2,
|
||
"cost_usd": round(estimate_cost(PRICED, 2000, 400), 6), "unpriced_calls": 0}
|
||
assert got["by_model"][UNPRICED]["unpriced_calls"] == 1 and got["by_model"][UNPRICED]["cost_usd"] == 0
|
||
# 目录外的模型名（多了后缀）照样算估不出；目录价是 0 的本地模型没报用量也不算估不出
|
||
assert got["by_model"][PRICED + "-silent"]["unpriced_calls"] == 1
|
||
assert got["by_model"][FREE]["unpriced_calls"] == 0
|
||
assert got["calls"] == 5 and got["unpriced_calls"] == 2
|
||
assert got["cost_usd"] == pytest.approx(estimate_cost(PRICED, 2000, 400))
|
||
|
||
|
||
async def test_outside_the_meter_nothing_is_counted():
|
||
m = usage.meter()
|
||
with usage.metering(m):
|
||
pass
|
||
await Scripted().ainvoke("你好")
|
||
assert m.summary()["calls"] == 0
|
||
assert usage.current() is None
|
||
|
||
|
||
async def test_metered_counts_the_generator_and_its_tasks_but_not_the_caller():
|
||
"""生成过程里的调用（含它起的任务）计进来；调用方在两次 yield 之间发起的「运行」不计。"""
|
||
m = usage.meter()
|
||
leaked: list[asyncio.Task] = []
|
||
|
||
async def generation():
|
||
await Scripted(model=PRICED).ainvoke("起草")
|
||
# 心跳那一层每一步起一个任务：任务在步里创建，拷走的上下文带着计量
|
||
await asyncio.ensure_future(Scripted(model=PRICED).ainvoke("挑表"))
|
||
yield "first"
|
||
await Scripted(model=PRICED).ainvoke("自查修正")
|
||
yield "second"
|
||
|
||
seen = []
|
||
async for item in usage.metered(generation(), m):
|
||
seen.append(item)
|
||
# 调用方拿到一条操作后自己调模型（相当于发起运行）：它的上下文里没有计量
|
||
await Scripted(model=UNPRICED).ainvoke("运行里的调用")
|
||
leaked.append(asyncio.ensure_future(Scripted(model=UNPRICED).ainvoke("运行任务")))
|
||
await asyncio.gather(*leaked)
|
||
got = m.summary()
|
||
assert seen == ["first", "second"]
|
||
assert got["by_model"] == {PRICED: got["by_model"][PRICED]} and got["calls"] == 3
|
||
assert usage.current() is None
|
||
|
||
|
||
async def test_nested_meters_fold_into_the_outer_one():
|
||
outer, inner = usage.meter(), usage.meter()
|
||
|
||
async def generation():
|
||
await Scripted(model=PRICED).ainvoke("起草")
|
||
yield 1
|
||
|
||
with usage.metering(outer):
|
||
await Scripted(model=PRICED).ainvoke("外层自己的")
|
||
async for _ in usage.metered(generation(), inner):
|
||
pass
|
||
with usage.metering(usage.meter()):
|
||
await Scripted(model=PRICED).ainvoke("复核")
|
||
assert inner.summary()["calls"] == 1
|
||
assert outer.summary()["calls"] == 3
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# answer_ops 末尾的 usage、复核的 usage
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
def _ops(*ops: dict) -> list[str]:
|
||
return [json.dumps(o, ensure_ascii=False) for o in ops]
|
||
|
||
|
||
GRAPH = _ops(
|
||
{"op": "plan", "summary": "数一数"},
|
||
{"op": "add_node", "node": {"id": "start", "type": "input", "label": "输入",
|
||
"config": {"fields": [{"name": "question"}]}}},
|
||
{"op": "add_node", "node": {"id": "lp", "type": "loop", "label": "循环",
|
||
"config": {"mode": "while", "condition": "foo(vars.x)", "max_iterations": 2}}},
|
||
{"op": "add_edge", "edge": {"source": "start", "target": "lp"}},
|
||
{"op": "add_node", "node": {"id": "body", "type": "transform", "label": "体",
|
||
"config": {"mode": "expression", "expression": "1"}}},
|
||
{"op": "add_edge", "edge": {"source": "lp", "target": "body", "sourceHandle": "body"}},
|
||
{"op": "add_edge", "edge": {"source": "body", "target": "lp"}},
|
||
{"op": "add_node", "node": {"id": "out", "type": "output", "label": "成果",
|
||
"config": {"fields": [{"name": "r", "value": "1"}]}}},
|
||
{"op": "add_edge", "edge": {"source": "lp", "target": "out", "sourceHandle": "done"}},
|
||
{"op": "done", "explanation": "数一数", "run": True},
|
||
)
|
||
FIX = _ops({"op": "update_node", "id": "lp", "config": {"mode": "while", "condition": "len(vars.x) > 1",
|
||
"max_iterations": 2}},
|
||
{"op": "done", "explanation": "改好了"})
|
||
|
||
|
||
@pytest.fixture
|
||
async def client():
|
||
async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
|
||
yield c
|
||
|
||
|
||
async def _sse(client, body: dict) -> list[dict]:
|
||
events = []
|
||
async with client.stream("POST", "/api/copilot/generate-stream", json=body) as r:
|
||
async for line in r.aiter_lines():
|
||
if line.startswith("data: "):
|
||
events.append(json.loads(line[6:]))
|
||
return events
|
||
|
||
|
||
async def test_answer_ops_ends_with_the_usage_of_every_build_call(client, monkeypatch):
|
||
"""起草、自查修正各一次。读到 done 那一行之后照样把模型的流读完（模型照约定 done 就是最后一行，读完只多收流末尾那一段
|
||
用量）：输出 token 报全了、金额估得出，这一轮的花费不再是下限（第 1 波的遗留：以前读到 done 就关流，输出 token 一个都
|
||
拿不到，两次都记成估不出）。"""
|
||
import app.api.copilot as copilot
|
||
|
||
model = Scripted(model=PRICED, replies=[GRAPH, FIX], tokens=(3000, 500))
|
||
|
||
async def _model(*_a, **_k):
|
||
return model, PRICED
|
||
|
||
monkeypatch.setattr(copilot, "get_chat_model", _model)
|
||
events = await _sse(client, {"instruction": "数一数", "intent": "answer"})
|
||
assert [e["op"] for e in events if e["op"] == "check"] and events[-2]["op"] == "final"
|
||
spent = events[-1]
|
||
assert spent["op"] == "usage", spent
|
||
assert model.calls == 2, spent # 起草一次、自查修正一次
|
||
row = spent["by_model"][PRICED]
|
||
assert row["calls"] == 2, spent
|
||
assert (row["input_tokens"], row["output_tokens"]) == (6000, 1000), "流末尾才报的输出 token 也收到了"
|
||
assert spent["unpriced_calls"] == 0, "读完了的流用量报全了，不记估不出"
|
||
assert spent["cost_usd"] == pytest.approx(estimate_cost(PRICED, 6000, 1000), abs=1e-6)
|
||
|
||
|
||
async def test_a_look_line_also_waits_for_the_usage_at_the_end_of_the_stream(client, monkeypatch):
|
||
"""读到 look 停下查看、ask 等用户回答：同样把这一次的流读完再关，查看前那一次的输出 token 不丢。"""
|
||
import app.api.copilot as copilot
|
||
|
||
looked = _ops({"op": "look", "kind": "table", "id": "不存在的表", "need": "看看有哪些列"})
|
||
model = Scripted(model=PRICED, replies=[looked, _ops({"op": "ask", "question": "要按哪一天算？"})], tokens=(1200, 80))
|
||
|
||
async def _model(*_a, **_k):
|
||
return model, PRICED
|
||
|
||
monkeypatch.setattr(copilot, "get_chat_model", _model)
|
||
events = await _sse(client, {"instruction": "看一下再问", "intent": "answer"})
|
||
spent = events[-1]
|
||
assert spent["op"] == "usage" and model.calls == 2, [e["op"] for e in events]
|
||
row = spent["by_model"][PRICED]
|
||
assert (row["input_tokens"], row["output_tokens"], row["unpriced_calls"]) == (2400, 160, 0), spent
|
||
|
||
|
||
class Chatty(Scripted):
|
||
"""done 那一行之后还往下写一大段闲话（不守约定），用量照样在最后一段才报。"""
|
||
|
||
chatter: int = 40
|
||
|
||
async def _astream(self, messages, stop=None, run_manager=None, **kwargs):
|
||
lines = self._next()
|
||
inp, out = self.tokens
|
||
for i, line in enumerate(lines):
|
||
meta = {"input_tokens": inp, "output_tokens": 0, "total_tokens": inp} if i == 0 else None
|
||
yield ChatGenerationChunk(message=AIMessageChunk(content=line + "\n", **({"usage_metadata": meta} if meta
|
||
else {})))
|
||
for _ in range(self.chatter):
|
||
yield ChatGenerationChunk(message=AIMessageChunk(content="这是一段写在最后一行之后、没人要的闲话。\n"))
|
||
yield ChatGenerationChunk(message=AIMessageChunk(content="", usage_metadata={
|
||
"input_tokens": 0, "output_tokens": out, "total_tokens": out}))
|
||
|
||
|
||
async def test_chatter_after_the_last_line_is_cut_off_and_recorded_as_not_fully_reported(client, monkeypatch):
|
||
"""读完流只为拿用量，不为闲话付钱：最后一行之后又收到超过上限的字就当场关流。这一次的用量没报全，照实记成估不出，不当成
|
||
报全了，也不悄悄当成 0 元。"""
|
||
import app.api.copilot as copilot
|
||
from app.api import copilot_gen
|
||
|
||
model = Chatty(model=PRICED, replies=[_ops({"op": "ask", "question": "要按哪一天算？"})], tokens=(500, 300))
|
||
|
||
async def _model(*_a, **_k):
|
||
return model, PRICED
|
||
|
||
monkeypatch.setattr(copilot, "get_chat_model", _model)
|
||
assert model.chatter * len("这是一段写在最后一行之后、没人要的闲话。\n") > copilot_gen.DRAIN_CHARS
|
||
events = await _sse(client, {"instruction": "按天看销售额", "intent": "answer"})
|
||
assert events[-1]["op"] == "usage" and model.calls == 1, [e["op"] for e in events]
|
||
row = events[-1]["by_model"][PRICED]
|
||
assert (row["calls"], row["input_tokens"], row["output_tokens"], row["unpriced_calls"]) == (1, 500, 0, 1), row
|
||
|
||
|
||
async def test_the_eval_driver_records_the_same_usage_the_chat_page_receives(client, monkeypatch):
|
||
"""评测的作答（app/eval/answer.Driver._generate → _take）收下的助手用量，就是问数据页经 SSE 收到的那条 usage：同一个
|
||
生成过程（answer_ops）、同一份计量，读完流之后两边都是报全了的数。"""
|
||
import app.api.copilot as copilot
|
||
from app.eval import answer
|
||
|
||
replies = [_ops({"op": "ask", "question": "要按哪一天算？"})]
|
||
model = Scripted(model=PRICED, replies=replies, tokens=(700, 90))
|
||
|
||
async def _model(*_a, **_k):
|
||
return model, PRICED
|
||
|
||
monkeypatch.setattr(copilot, "get_chat_model", _model)
|
||
chat = (await _sse(client, {"instruction": "按天看销售额", "intent": "answer"}))[-1]
|
||
assert chat["op"] == "usage"
|
||
driver = answer.Driver("按天看销售额", datasource_ids=[], ctx=answer.TurnContext(
|
||
suite_id="s", suite_name="虚构题集", eval_run_id="e", case_id=None, case_seq=1, actor=None))
|
||
ops, refused = await driver._generate("按天看销售额", None)
|
||
assert refused is None and any(o.get("op") == "ask" for o in ops)
|
||
driver._take(ops)
|
||
got = driver.turn.usage["assistant"]
|
||
assert got["by_model"] == chat["by_model"] and got["cost_usd"] == chat["cost_usd"], (got, chat)
|
||
assert got["by_model"][PRICED]["output_tokens"] == 90 and got["unpriced_calls"] == 0
|
||
|
||
|
||
async def test_a_build_on_an_unpriced_model_says_it_cannot_be_priced(client, monkeypatch):
|
||
import app.api.copilot as copilot
|
||
|
||
model = Scripted(model=UNPRICED, replies=[_ops({"op": "reply", "text": "已经查过了"})])
|
||
|
||
async def _model(*_a, **_k):
|
||
return model, UNPRICED
|
||
|
||
monkeypatch.setattr(copilot, "get_chat_model", _model)
|
||
events = []
|
||
async with client.stream("POST", "/api/copilot/generate-stream", json={"instruction": "刚才那个数是多少"}) as r:
|
||
async for line in r.aiter_lines():
|
||
if line.startswith("data: "):
|
||
events.append(json.loads(line[6:]))
|
||
assert [e["op"] for e in events] == ["model", "reply", "usage"]
|
||
assert events[-1]["calls"] == 1 and events[-1]["unpriced_calls"] == 1 and events[-1]["cost_usd"] == 0
|
||
|
||
|
||
async def test_a_stream_read_to_the_end_is_fully_priced():
|
||
"""读完的流（直接回答那一行之后模型自己停了）：输入、输出 token 都报全了，金额估得出。"""
|
||
m = usage.meter()
|
||
|
||
async def generation():
|
||
async for _ in Scripted(model=PRICED, tokens=(400, 60)).astream("问"):
|
||
pass
|
||
yield 1
|
||
|
||
async for _ in usage.metered(generation(), m):
|
||
pass
|
||
row = m.summary()["by_model"][PRICED]
|
||
assert (row["input_tokens"], row["output_tokens"], row["unpriced_calls"]) == (400, 60, 0)
|
||
assert row["cost_usd"] == round(estimate_cost(PRICED, 400, 60), 6)
|
||
|
||
|
||
async def test_the_review_carries_its_own_usage(client, monkeypatch):
|
||
import app.api.copilot as copilot
|
||
from app.core.events import EventType
|
||
from app.db.base import SessionLocal
|
||
from app.db.models import Run, RunEvent
|
||
|
||
reply = json.dumps({"note": "agent 用满了步数，结论不完整。", "answer": "", "retry": True}, ensure_ascii=False)
|
||
model = Scripted(model=PRICED, replies=[[reply]], tokens=(800, 100))
|
||
|
||
async def _model(*_a, **_k):
|
||
return model, PRICED
|
||
|
||
monkeypatch.setattr(copilot, "get_chat_model", _model)
|
||
async with SessionLocal() as session:
|
||
run = Run(status="succeeded", output={"answer": "大概是 3 家"}, graph={})
|
||
session.add(run)
|
||
await session.flush()
|
||
session.add(RunEvent(run_id=run.id, seq=1, type=EventType.LOG, node_id="a", ts=0.0,
|
||
data={"level": "warn", "code": "step_limit", "message": "用满了步数"}))
|
||
await session.commit()
|
||
run_id = run.id
|
||
body = (await client.post("/api/copilot/review", json={"run_id": run_id, "question": "有几家门店"})).json()
|
||
assert body["usage"]["calls"] == 1 and body["usage"]["by_model"][PRICED]["input_tokens"] == 800
|
||
# 干净的运行不调模型：usage 全 0
|
||
async with SessionLocal() as session:
|
||
clean = Run(status="succeeded", output={"answer": "3 家"}, graph={})
|
||
session.add(clean)
|
||
await session.commit()
|
||
clean_id = clean.id
|
||
body = (await client.post("/api/copilot/review", json={"run_id": clean_id, "question": "有几家门店"})).json()
|
||
assert body["verdict"] == "ok" and body["usage"]["calls"] == 0 and body["usage"]["by_model"] == {}
|